From a3b33e4103e0a90112a730d249635c2c766e57d3 Mon Sep 17 00:00:00 2001
From: Giulio Eulisse <10544+ktf@users.noreply.github.com>
Date: Tue, 22 Sep 2026 13:35:51 +0200
Subject: [PATCH] MathUtils: make SMatrixGPU compile as MSL

Three separate issues:

* Metal has no static storage duration inside a function, so
  MatRepSymGPU::off() drops the static on its offset table. Nothing is
  lost by dropping it everywhere rather than forking: arrays, unlike
  scalars, are emitted into read-only data either way. On the host it is
  marginally better, since the static form reached the table through the
  GOT.

* operator=(Expr) now takes the expression by value. MSL folds the
  address space into a deduced parameter, so the generic
  operator=(const M&) deduced M = thread Expr<...>, matched exactly, beat
  the Expr overload and then tried mRep = rhs.mRep on an expression. A
  by-value parameter has no address space to deduce, so the Expr overload
  wins on Metal as it already did elsewhere. Expr is two references, so
  there is no copy cost.

* Metal supports up to C++17, so it needs the same treatment as OpenCL
  for a requires-clause.
---
 Common/MathUtils/include/MathUtils/SMatrixGPU.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/Common/MathUtils/include/MathUtils/SMatrixGPU.h b/Common/MathUtils/include/MathUtils/SMatrixGPU.h
index 8158a93666a92..2e551aac98d24 100644
--- a/Common/MathUtils/include/MathUtils/SMatrixGPU.h
+++ b/Common/MathUtils/include/MathUtils/SMatrixGPU.h
@@ -340,7 +340,7 @@ class MatRepSymGPU
 
   static GPUdi() int off(int i)
   {
-    static constexpr auto v = row_offsets_utils::make<D * D>(off1);
+    constexpr auto v = row_offsets_utils::make<D * D>(off1);
     return v[i];
   }
 
@@ -453,7 +453,7 @@ class SMatrixGPU
   GPUd() SMatrixGPU(const SMatrixGPU<T, D1, D2, R2>& rhs);
   template <class A, class R2>
   GPUd() SMatrixGPU(const Expr<A, T, D1, D2, R2>& rhs);
-  template <class M>
+  template <class M, class = decltype(M::mRep)>
   GPUd() SMatrixGPU<T, D1, D2, R>& operator=(const M& rhs);
   template <class A, class R2>
   GPUd() SMatrixGPU<T, D1, D2, R>& operator=(const Expr<A, T, D1, D2, R2>& rhs);
@@ -518,7 +518,7 @@ class SMatrixGPU
   R mRep;
 };
 
-#ifndef __OPENCL__ // TODO: current C++ for OpenCL 2021 is at C++17, so no concepts. But we don't need this trick for OpenCL anyway, so we can just hide it.
+#if !defined(__OPENCL__) && !defined(__METAL__) // TODO: current C++ for OpenCL 2021 and MSL 4.1 are both at C++17, so no concepts. But we don't need this trick there anyway, so we can just hide it.
 template <class T, unsigned int D1, unsigned int D2, class R, typename Y, typename X = Y>
   requires(sizeof(typename X::traits_type::pos_type) != 0) // do not provide a template to fair::Logger, etc... (pos_type is a member type of all std::ostream classes)
 GPUd() X& operator<<(Y& y, const SMatrixGPU<T, D1, D2, R>&)
@@ -684,8 +684,8 @@ GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator=(const Expr
 }
 
 template <class T, unsigned int D1, unsigned int D2, class R>
-template <class M>
-GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator=(const M & rhs)
+template <class M, class>
+GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator=(const M& rhs)
 {
   mRep = rhs.mRep;
   return *this;